The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The architecture documentation describes a process where API keys (Anthropic, Gemini, Flux2) are discovered from environment variables or config files and then injected into a {PROJECT_CODE}_data.js file. This file is intended to be loaded by a client-side browser dashboard, effectively exposing the credentials to any user or environment with access to the dashboard.
[COMMAND_EXECUTION]: The skill contains logic to generate, copy, and write files to the project directory, including HTML templates, JavaScript data payloads, and log files in the 'AI
Project Brain/' folder.
[DATA_EXFILTRATION]: The skill enables direct API calls from the browser to api.anthropic.com using the anthropic-dangerous-direct-browser-access header. This pattern, combined with client-side key exposure, allows for potential unauthorized access to the developer's API accounts.
[PROMPT_INJECTION]: The AI chatbot component processes external project data (RFIs, submittals, daily reports) as context for answering queries. This creates a surface for indirect prompt injection attacks where malicious actors could embed instructions in project documents to manipulate the AI's behavior.
Ingestion points: submittals.json, rfis.json, daily-reports.json, meetings.json, change-orders.json, inspections.json (as described in SKILL.md and references/skill-detail.md).
Boundary markers: None identified in the provided chatbotConfig system prompt logic.
Capability inventory: File writing ({PROJECT_CODE}_data.js, {PROJECT_CODE}_Data_Intel.html) and outbound network requests to Anthropic API.
Sanitization: The skill utilizes DOMPurify for UI rendering but does not specify sanitization for data interpolated into LLM prompts.
[EXTERNAL_DOWNLOADS]: The dashboard template references multiple third-party libraries (Chart.js, Fabric.js, D3.js, DOMPurify) via the cdn.jsdelivr.net CDN. These are well-known technology services.

project-data-intel